8-2 幾何圖神經網絡在藥物發現中的應用.pdf

1、?B?.?/?.?/?:?/?/.?CONTENTSCONTENTS3?1 1?2 23 3?01?PART?e?d?b?:?ea?Ic:?Q?V?2?-?:-?.?./?.2?.?0?.2?2?A?R?A?2?D?7F?0?4?26?1?26?26?26?https:/ to model 3D spatial structure of the complex effectively?-Molecular graph information should be preserved-Spatial position should be independent of cartesian coord

2、inatesyzxxyzRotationCoordinate system 1Coordinate system 2Inconsistent spatial information for model learning?How to model 3D spatial structure of the complex effectively?-Molecular graph information should be preserved-Spatial position should be independent of cartesian coordinatesyzxxyzRotationCoo

3、rdinate system 1Coordinate system 2Model should capture the relative spatial position in the complexDistance and Angle?Solutions-Equivariant neural networks-Tensor field networks:Rotation-and translation-equivariant neural networks for 3d point clouds,2018-Se(3)-transformers:3d roto-translation equi

4、variant attention networks,NeurIPS 2020,-E(n)equivariant graph neural networks,ICML 2021-Geometric encoded message passing-Schneta deep learning architecture for molecules and materials,JCP,2018-PhysNet:A neural network for predicting energies,forces,dipole moments,and partial charges,JCTCCE,2019-Di

5、mNet:Directional message passing for molecular graphs,ICLR 2020-Equivariance02?PART?3D?Virtual Screening Aiming to screen out the drugs which can bind with the target protein The process is time-consuming and requires expert knowledge The prediction of binding affinity can accelerate this processStr

6、ucture-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity“;KDD 2021?Protein-Ligand Binding Affinity Binding affinity is defined as the strength of the binding interaction between a protein(target)and a ligand(drug)It can be used to rank drug candidates in d

7、rug discovery Protein and drug molecules have 3-dimensional structures naturallyproteinligandStructure-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity“;KDD 2021 Structure-based Binding Affinity Prediction Focusing on learning from 3D-structure protein-li

8、gand complexes 3D structural information can effectively contribute to the drug design Traditional data-driven methods Failing to combine 3D structural and topological information Ignoring the molecule-level interaction?Structure-aware Interactive Graph Neural Networks for the Prediction of Protein-

9、Ligand Binding Affinity“;KDD 2021 We apply a sampling-based process to construct the directional interaction graph based on 3D atomic coordinates Preserve nodes within!Distance factor All ligands atoms and partial proteins atoms close to ligand Preserve edges within!Angle factor Local covalent bonds

10、 Non-local correlations Complex Interaction Graph ConstructionPreserving the key structure of the complex?Structure-aware Interactive Graph Neural Network(SIGN)Structure-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity“;KDD 2021 Establish a local polar co

11、ordinate system in GNN Preserve angle informationAngle domain division Preserve distance informationDistance discretization Apply a nodeedgeinteraction schema Polar Coordinate-Inspired Graph Attention01 distance12 distance23 distance3 Distance discretizationAngle domain division?Angle-oriented NodeE

12、dge Interaction Aggregating the node features for edge Dividing angle domains around L-axis Polar Coordinate-Inspired Graph AttentionAggregation for edge#$NodeEdge Interaction LayerAngle domain index of neighbor%&Edge-oriented neighbors in(-)angle domain?Angle-oriented NodeEdge Interaction Domain-sp

13、ecific aggregation process Global aggregation for edge embedding Polar Coordinate-Inspired Graph AttentionAggregation for edge#$NodeEdge Interaction Layer?Distance-aware EdgeNode Interaction Discriminate multiple spatial relations in attention Polar Coordinate-Inspired Graph AttentionAggregation for

14、 node#EdgeNode Interaction Layer$#%$&%$()%*#%*&%Distance injectionAngle injection?PDBbind 3D binding structures of protein-ligand complexes and experimental affinities Three overlapping subsets CSAR-HiQ An additional benchmark dataset(104 complexes used)?Datasetscorerefinedgeneralcore set:290 comple

15、xes=testingrefined set:4,057 complexes=traininggeneral set:13,283 complexes=training Baselines ML-LR,SVR,RF-Score CNN-Pafnucy,OnionNet Seq-GNN-GraphDTA:GCN,GAT,GIN Spa-GNN-SGCN,GNN-DTI,DMPNN,MAT,CMPNN,DimeNet Comparison with baselinesOur proposed model SIGN achieves the best performance on two bench

16、marks?Structure-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity“;KDD 2021 Validate the effectiveness of distance and angle learning Removal of both spatial factors(SIGN-AD)One-sided spatial structural information(SIGN-D and SIGN-A)Validate the effectiven

17、ess of using long-range interactions Removal of interaction loss(SIGN-I)Impact of Spatial and Interactive FactorsContribution of spatial and interactive factors?Structure-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity“;KDD 2021 3D structure and inter-mo

18、lecule interaction are both essential for thestructure-based protein-ligand binding affinity prediction.The polar-inspired graph attention is designed to integrate both distance and angle information for 3D spatial structure modeling.We proposed a well-designed pooling process along with a reconstru

19、ction learning task for pairwise interaction matrix.Experimental results on two benchmarks showed the effectiveness and the generalizability of the proposed model.Summary of our work?Structure-aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affinity“;KDD 2021?Mol

20、ecular Property Prediction Learning the molecular representation by graph methods Aiming to predict the target property based on the representation Helping to accelerate the biological and chemical-related process?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2

21、022?Graph Representation Learning for Moleculesthe scarcity of labeled data?the unique geometric characteristicsSelf-supervised LearningGeometry-based Graph Learning?Most previous works adopt the vanilla GNN model Molecular graph is different from the general graph?GeomGCL:Geometric Graph Contrastiv

22、e Learning for Molecular Property Prediction?;AAAI 2022?Geometric Structure Learning on Graphs Utilize relative geometry factors,such as distance and angle Almost focus on single-view geometric learning?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?Contras

23、tive Learning on Graphs Utilize multiple molecular graphs for self-supervision Classic data augmentations can alter the molecular structures?Contrastive Learning on Graphs Utilize multiple molecular graphs for self-supervision Classic data augmentations can alter the molecular structures Our method-

24、leverage the geometric learning and contrastive learning?Overall Framework for GeomGCL Geometric encoding based on dual-view molecular graphs construction Message passing neural network with geometry-enhanced learning 2D-3D molecular graph contrastive optimization?Geometry-based RBF Encoding 2D view

25、 graph:local distance!and 2D angle 3D view graph:global distance#and 3D angle$Radial Basis Function Kernel%&()&(*&+&?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?Adaptive Geometric Message Passing Scheme Graph-level representations for 2D and 3D views Geo

26、metry-enhanced contrastive learning Spatial regularized constraint!#!$#%#%$#&()*+,)-./1234+)-+5Loss2D ProjectionHead3D ProjectionHead+?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?Dataset Seven benchmark datasets from MoleculeNet Baselines Message Passing

27、 Methods AttentiveFP,CoMPT,DMPNN Geometry-based GNNs SGCN,MAT,HMGNN,DimeNet GCL Methods InforGraph,MoCL?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?Comparison with baselines One-side geometric information is not effective enough?GeomGCL:Geometric Graph C

28、ontrastive Learning for Molecular Property Prediction?;AAAI 2022?Comparison with baselines One-side geometric information is not effective enough Domain knowledge is helpful for contrastive learning?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?Comparison

29、with baselines One-side geometric information is not effective enough Domain knowledge is helpful for contrastive learning Our proposed methods achieve the best performance on seven benchmarks?The effectiveness of 2D-3D geometric information Dual-view geometric learning enhances the molecular repres

30、entation 2D-3D geometric contrastive strategy is effective The effect of coefficient for spatial regularizer loss?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?View Visualization The self-supervised method GeomGCL learns useful information?GeomGCL:Geometri

31、c Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?Summary of our work Design the dual-channel geometric MPNN to capture the distance and angle information under both 2D and 3D views Builds the bridge between the geometric structure learning and the graph contrastive learning

32、The experimental results on the downstream property prediction tasks demonstrate the effectiveness of the proposed GeomGCL?GeomGCL:Geometric Graph Contrastive Learning for Molecular Property Prediction?;AAAI 2022?E?m?e?h?b?ca?4?M?E?k?ca?e?G1?Atom-bond graphBond-angle graph3D spatial structureCNH1H2H

33、3H4H5Atom:Bond:Angle:?Geometry Enhanced Molecular Representation Learning for Property Prediction?;Nature Machine Intelligence 2022?%?b?I?V?8H?ti4?V?2?.1?.1?2?4?bx?cM?E?o?a?4?G?s?ln?e?o?a?ln?e?V?I?t?cM?%Geometry Enhanced Molecular Representation Learning for Property Prediction?;Nature Machine Intelligence 202204?PART?c?TG?r?M b?I?t?9?-?-?:?r?M?s?h?e?h?g?PI?e?mAD?ua?ep?Id?ih?x?Hlno?Em?/?.?/9?-?-?-?-?:/?-/-?:?9?9.?-?cn?/?.?n?n?:?d?ab?n?.?2022?!